Real-ESRGAN
Open-source image and video super-resolution and restoration models trained on pure synthetic data, with Python and portable ncnn-Vulkan inference.
At a Glance
Real-ESRGAN core software released under the OSI-approved BSD 3-Clause license; free to use, modify and redistribute under license terms.
Engagement
Available On
Alternatives
Listed Nov 2026
About Real-ESRGAN
Real-ESRGAN is an open-source project by Xintao Wang and collaborators (Tencent ARC Lab and the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences) for practical image and video restoration. It extends ESRGAN into a blind super-resolution approach trained with pure synthetic data. The repository provides pre-trained models, a Python inference script, training code, and portable executables for Windows, Linux and macOS.
What It Is
Real-ESRGAN is a super-resolution and restoration toolkit built on PyTorch. Users feed in low-quality or low-resolution images, or anime video, and get upscaled, restored output. The README describes the aim as developing practical algorithms for general image and video restoration.
Models and Inference Options
Several pre-trained models are documented: RealESRGAN_x4plus for general images, RealESRGAN_x4plus_anime_6B for anime illustrations, RealESRGAN_x2plus, the AnimeVideo-v3 model for anime video, and the small realesr-general-x4v3 model, which has a -dn option to balance denoising strength. The Python script supports tiling, alpha-channel, grayscale and 16-bit images, arbitrary output scales via --outscale, and optional GFPGAN face enhancement. Fine-tuning on your own data is supported, and training code is released.
Ways to Run It
There are three documented routes: online demos (Replicate, Colab, Hugging Face Spaces), portable ncnn-Vulkan executables for Intel, AMD and Nvidia GPUs that need no CUDA or PyTorch, and the Python script, which requires Python 3.7 or later and PyTorch 1.7 or later. The executable processes images in tiles, so it can show block inconsistency and differ slightly from the PyTorch results, and it does not support every option of the Python script, such as outscale.
Ecosystem
The README lists community projects built on Real-ESRGAN, including a VapourSynth plugin, an Android ncnn port and several GUI front ends such as Upscayl.
Community Discussions
Be the first to start a conversation about Real-ESRGAN
Share your experience with Real-ESRGAN, ask questions, or help others learn from your insights.
Pricing
Open Source (BSD-3-Clause)
Real-ESRGAN core software released under the OSI-approved BSD 3-Clause license; free to use, modify and redistribute under license terms.
- BSD 3-Clause License
- Python package on PyPI (realesrgan)
- Portable ncnn-vulkan executables for Windows/Linux/MacOS (Intel/AMD/Nvidia GPU)
- Training and inference code and model zoo
- Hosted demos (Replicate, Colab) are separate from the licensed core
Capabilities
Key Features
- Blind super-resolution trained on pure synthetic data
- General image model RealESRGAN_x4plus
- Anime image model RealESRGAN_x4plus_anime_6B
- Anime video model AnimeVideo-v3
- Small general model realesr-general-x4v3 with denoise strength option
- Arbitrary output scale via --outscale
- Tile option for large images
- Supports alpha channel, grayscale and 16-bit images
- Optional GFPGAN face enhancement
- Portable ncnn-Vulkan executables for Windows, Linux and macOS
- Fine-tuning and training on custom datasets
Integrations
Demo Video

